138results about How to "Avoid vanishing gradients" patented technology

The invention relates to a marine ranch power supply system line fault diagnosis method based on MS-2DResNet and ICBAM, and belongs to the field of power supply system line fault diagnosis. The fault diagnosis method is mainly divided into two stages, namely, multi-scale feature extraction and attention mechanism feature fusion. Firstly, a preprocessed fault data sample is converted into two-dimensional data to be input into a fault diagnosis model; in the first stage, local and global feature extraction is performed on the converted fault data by using multi-scale convolution, so that the robustness of the model for fault data features of different sizes and scales is improved. And then, entering an attention feature fusion stage, and carrying out weighted fusion on stacked data from different scale channels by utilizing the improved CBAM, so that the model pays more attention to useful fault features and neglects useless features, and further, the fault diagnosis accuracy of the model is improved. The method has the characteristics of high diagnosis accuracy and good anti-noise performance.

The invention relates to the technical field of novel power system optimization operation, in particular to a new energy high-uncertainty scene generation method and device based on a CWGAN-GP model, and the method comprises the steps: correcting an initial generative adversarial network through employing a Wasserstein distance and a gradient penalty term, so as to obtain a final generative adversarial network, and the target convolutional neural network is introduced into the final generative adversarial network to obtain a conditional deep convolutional generative adversarial network of improved deep convolution, so that WGAN-GP is introduced into the conditional deep convolutional generative adversarial network of improved deep convolution to obtain a CWGAN-GP model of improved deep convolution. And a target L-ISODATA clustering algorithm is applied to cluster the generated wind and light scenes so as to generate a high-uncertainty new energy power generation output scene. Therefore, the problems that in the related technology, when the GAN processes high-dimensional data, complex space-time characteristics are difficult to effectively capture, and due to lack of prior information, the training direction is prone to deviation, and the scene generation effect and stability are affected are solved.

The invention discloses a coordinated optimization multi-step prediction method based on target correlation space-time coding-task decoding, and the method comprises the steps: constructing a space-time encoder formed by stacking a plurality of space attention enhanced time sequence convolution modules, and carrying out the complete adaptive noise set empirical mode decomposition of a target value historical sequence; constructing a parameter sharing mechanism, and training a multi-step prediction model formed by a space-time encoder and a task specific decoder by using a multi-task learning normal form; training input is preprocessed historical industrial process production data, and gradient conflicts among multiple tasks are reduced by using a gradient adjustment mechanism and a dynamic weight distribution method; and preprocessing real-time industrial process production data, and inputting the data into the trained prediction model to obtain a prediction result. According to the method, the relationship among different prediction tasks can be effectively coordinated, the overall model performance is optimized, and challenges encountered in long-term prediction are more effectively processed by dynamically adjusting weight distribution and gradient flow among multiple tasks.

The invention discloses a project management system and method based on a generative adversarial network. The system comprises a data acquisition unit, a feature extraction unit, a progress prediction unit and a scheme generation unit. Wherein the feature extraction unit is used for respectively extracting time sequence features and space features by adopting a long-short term memory network and a multi-head image attention network; the progress prediction unit generates a prediction sequence through a generator based on a generative adversarial network architecture, and evaluates the rationality of the prediction sequence through a discriminator; the scheme generation unit generates a progress adjustment scheme based on key path analysis, resource optimization and a buffer time allocation method; according to the invention, the accuracy and reliability of project progress prediction can be effectively improved, and a scientific basis is provided for project management decision.

The invention belongs to the technical field of computer application, and relates to an operation inspection equipment fault detection method and system based on an improved deep learning model, and the method comprises the steps: obtaining original data, employing EEMD to decompose the original data for feature enhancement, obtaining a decomposition signal, carrying out the noise reduction of the decomposition signal, and obtaining a feature signal after noise reduction; according to the feature signals after noise reduction, a deep learning model is constructed based on Resnet and a dynamic Transform encoder, and the prediction probability of each fault category is output; a ZOA optimization algorithm is used to optimize hyper-parameters of a deep learning model, an optimization model is obtained, the performance of the optimization model is evaluated, an optimal model is obtained, and the prediction probability of each fault category is output, so that false report and missing report of faults can be effectively reduced in a fault detection task, potential safety hazards caused by the fact that the faults are not found in time are reduced, and the fault detection efficiency is improved. Therefore, the stability and reliability of the power system are improved.

The invention provides an OFDM signal receiving and processing system and method suitable for an NAVDAT system. According to the method, the characteristics that NAVDAT system signals have synchronous headers, pilot frequencies and the like are fully considered, firstly, the synchronous headers are used for carrying out correlation operation on received data to search a frame starting point, and a signal frame to be demodulated and decoded is obtained in combination with a constant false alarm detection algorithm; and subsequently, carrying out symbol synchronization to determine the FFT window position. Secondly, in combination with a discrete pilot signal, estimating an integral multiple frequency offset error, namely coarse synchronization, and further realizing estimation of a decimal multiple frequency offset error, namely fine synchronization, so as to realize frequency synchronization. And then, channel estimation is carried out by using the pilot signal and adopting a least square method to obtain the channel frequency response of the pilot, the frequency response of the complete signal is obtained through interpolation, and the rough estimation result is further processed through the designed RCAFCN network to obtain accurate channel estimation, so that the original OFDM signal is recovered. And finally, extracting a data part in the OFDM signal according to a frame structure, realizing soft demodulation through a log-likelihood ratio mode, decoding data of a signaling part by using a polarization code, and decoding data of a service part by using a low-density parity check code, thereby realizing complete demodulation and decoding. The method provided by the invention can effectively reduce inter-symbol interference and inter-carrier interference existing in OFDM signals, compensate channel distortion, reduce noise influence, improve decoding precision, improve communication quality, realize optimization of NAVDAT system baseband processing, and guarantee sailing safety of ships on the sea by combining with an existing marine broadcast system.

The invention discloses a human body behavior recognition method and system based on space-time decoupling multi-scale feature fusion. The system comprises a local space-time multi-scale feature fusion network MDCNN and a space-time decoupling network DTST. Extracting different scale features of multichannel sensor data by using a multi-scale convolutional neural network, extracting shallow layer features and deep layer features of signals at the same time, and extracting features representing action information by fully utilizing complementary information among different level features; the method comprises the following steps of: decoupling a network into a time decoupling structure and a space decoupling structure through a time-space decoupling network based on Transform, so as to accurately extract time-space characteristics in an inertial signal; and finally, using a Softmax classifier to carry out action classification and recognition. According to the method, consistency and complementarity among multi-sensor data can be efficiently utilized, multi-scale spatial-temporal characteristics are fused, changes in time dimensions and relevant information in spatial positions are fully considered, different attention degrees are kept for signals of different sensing channels, and the recognition precision in human behavior recognition is improved.

The invention provides a defect classification method and system for a semiconductor chip and a readable storage medium. The training method comprises the following steps: acquiring training set data, wherein the training set data comprises different types of defect images of the semiconductor chip; inputting the training set data into an initial defect classification model based on a deep convolutional neural network so as to train the initial defect classification model; in the training process, an attention mechanism module is introduced to the tail of at least part of convolution blocks of the initial defect classification model, and residual connection is introduced to at least part of the convolution blocks; and a target defect classification model is obtained after multiple rounds of iterative training. The method effectively balances model performance and computing resource requirements, and ensures that higher efficiency is kept while the classification precision is improved.

The invention discloses a three-dimensional human body posture estimation method and system based on machine learning. The method comprises the following steps: constructing a posture estimation data set; training and optimizing network parameters of the feature extraction model, the feature dimension reduction model and the regression model; and extracting a feature vector of the target image by using the feature extraction model, performing feature dimension reduction by using the feature dimension reduction model, and finally inputting the feature vector after dimension reduction into the regression model to predict the three-dimensional coordinates of each key point of the human body. According to the method, the generative adversarial network based on two-stage enhancement is used for data expansion, so that the generalization ability of the model can be improved; network parameters of the feature extraction model are optimized by adopting a neural network parameter optimization algorithm based on a fusion phenomenon, so that the problem of gradient disappearance or explosion is avoided; a sparse self-expression layer is introduced between an encoder and a decoder, so that the correlation between features is reduced; and the prediction strategy of the regression model is adjusted by adopting a reverse reinforcement learning mechanism, so that the prediction accuracy is remarkably improved.

The invention discloses a sintering end point multi-step prediction method based on target correlation space-time coding-task decoding. The method comprises the following steps: constructing a space-time encoder formed by stacking a plurality of space attention enhanced time sequence convolution modules; completely adaptive noise set empirical mode decomposition is carried out on a target value historical sequence, a screened internal mode function is input into a gated neural network and then passes through a full connection layer, and a result is input into a task specific decoder constructed by adopting a time attention mechanism; historical industrial process production data is preprocessed to train a prediction model composed of a space-time encoder and a task-specific decoder; and preprocessing real-time industrial process production data, and inputting the data into the trained prediction model to obtain a prediction result. According to the method, features can be effectively extracted from multiple levels, implicit information in historical data of target variables can be deeply mined, and the prediction capability of the model on the sintering end-point state is remarkably enhanced.

The invention discloses a photovoltaic power station supervised data generation method based on a generative adversarial network, which is used in the technical field of data augmentation, and comprises the following steps: initializing a generator and a discriminator, and randomly extracting training samples; training a generator and a discriminator by using the training sample, and generating a synthetic data set conforming to photovoltaic power station data set distribution through the generator; and clustering the features of the synthetic data set and the photovoltaic power station data set through a K-means clustering algorithm, observing a clustering center and feature distribution, evaluating the feature distribution difference between the generated data set and the real photovoltaic power station data set through an unsupervised clustering method, and evaluating the quality of the generated data set. According to the method, through the steps of density value sorting, initial center selection and the like, fine division of clustering is achieved, and initial center selection deviation in the clustering process is effectively avoided by combining density value calculation and a grid division method.

The invention discloses a chromosome segmentation method, system and device based on chromosome clusters and a medium, and belongs to the technical field of biomedicine and artificial intelligence in bioinformatics, and the segmentation method comprises the steps: carrying out the labeling processing of a chromosome original image, sequentially constructing a chromosome cluster recognition data set for different tasks, and training a chromosome cluster recognition network; constructing a chromosome cluster classification data set, and training a chromosome cluster classification network; constructing a chromosome cluster segmentation data set, and training a chromosome cluster segmentation network; inputting a chromosome original image, and performing reasoning by using the trained chromosome cluster recognition network, chromosome cluster classification network and chromosome cluster segmentation network in sequence to obtain a chromosome segmentation result; according to the invention, the chromosome can be automatically and accurately identified and segmented from the observation image; and by using the architecture based on the chromosome cluster, the probability of missing detection of the chromosome is reduced, and the downward compatibility of the network to the image resolution is high.

The invention relates to the field of image recognition or understanding, in particular to a face and tongue coat recognition and tea drink recommendation method and system based on artificial intelligence. The method comprises the following steps: identifying a plurality of image annotation labels, training a feature extraction model and a classification model by using the annotated images, obtaining a new image, performing feature extraction by using the trained feature extraction model, and inputting the feature into the trained classification model to obtain a label of the new image. The device is used for applying the method. The existing method and system for identifying the face and the tongue coating of the tea making machine and recommending the tea making have the problem that the identification of a face image and a tongue coating image is not accurate enough. The method for identifying the face and the tongue coating of the tea making machine and recommending the tea making based on the artificial intelligence provided by the invention is accurate in identification of the face image and the tongue coating image.

The invention provides a multi-modal agricultural disease and pest prediction method. The method comprises the following steps: acquiring image data and audio data by using a sensor, respectively preprocessing the image data and the audio data, extracting semantic features of the image data by using a CNN model, extracting MFCC features of the audio data by using a Mel system, and splicing and fusing the semantic features and the MFCC features to obtain fused features; the fusion features are input into a BiLSTM model according to time steps, an output result is obtained through training, the output result is input into a full connection layer for classification, probability distribution of all classes is obtained, a pest and disease damage prediction model is obtained, and a final model is obtained after verification set testing and parameter adjustment and used for prediction. Through feature extraction and fusion of the image and audio data, the BiLSTM model is used to carry out classification prediction on the fusion features, automatic prediction of agricultural diseases and insect pests is realized, and the prediction accuracy is higher on the basis of reducing manpower and time costs.

The invention discloses a data fusion method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: carrying out the classification of various types of sensor data in an unmanned system according to a set classification rule, and obtaining the classification data; performing preprocessing based on the basic attributes of the classified data to obtain preprocessed classified data; performing data-level fusion on the preprocessed classification data to obtain a plurality of classification fusion data; based on a deep recurrent neural network DRNN, in combination with a bidirectional long-short term memory unit Bi-LSTM and a sparse attention mechanism, constructing a data fusion model, and optimizing the data fusion model through model pruning to obtain an optimized data fusion model; and inputting the plurality of classified fusion data into the optimized data fusion model for secondary fusion to obtain a final fusion feature. The potential of multi-source information is fully mined, and the accuracy and effectiveness of decision making in a complex environment are improved.

The invention discloses a flaw detection method and device based on FabricMosaic and an improved fine-grained feature extraction network, relates to the technical field of cloth flaw detection, and solves the problems that in the prior art, a flaw target is incomplete and small target flaws are prone to missing detection due to the fact that a fixed cutting mode is adopted. A dynamic FabricMosaic data enhancement strategy based on target foreground and cloth texture background features is adopted to carry out sample expansion, images in a training set serve as input of a YOLOv5s network model, feature extraction is carried out through a feature fine-grained enhancement extraction module based on a divergent path to obtain a global fine-grained feature map, the YOLOv5s network model is trained, and a target foreground and cloth texture background feature is obtained. The final YOLOv5s network model is used for cloth flaw detection, a dynamic FabricMosaic data enhancement strategy is used for sample expansion, the integrity of a flaw target can still be effectively reserved in the sample integration process, and the risk of missing detection during small target flaw detection is reduced.

The invention belongs to the technical field of mobile phone display screen detection, and particularly discloses a defect detection method based on a smart mobile phone display screen. Graying and denoising the collected surface defect image of the original display screen; performing feature labeling and expansion on the preprocessed image; on the basis of the U-Net network, a residual block is inserted, common convolution is replaced by depth separable convolution, and an attention module is added; performing batch training on the cross entropy loss function by using a stochastic gradient descent method, continuously updating parameters of the network model, and optimizing the parameters of the network model by using an Adam method; when the output value of the loss function is reduced and does not change any more, iteration is stopped; and inputting a to-be-detected mobile phone display screen image into the trained network model for detection, if the mobile phone display screen is detected to have a defect, outputting a defect binary image of the mobile phone display screen, and calculating the real size of the defect. According to the invention, surface defects such as scratches and black spots of the mobile phone display screen can be rapidly and accurately detected.

The invention discloses a satellite image super-resolution method based on deep learning, which belongs to the field of image processing. The present invention builds an extensible deep super-resolution network, performs 4 times super-resolution on remote sensing satellite images, and can well reconstruct the characteristics of transmission line tower insulators, with the characteristics of high fidelity and fast processing speed; in order to improve the ability of the super-resolution network to learn the characteristics of remote sensing satellite images, further, a data augmentation method is adopted, and the weight normalization layer replaces the batch normalization layer to more accurately learn the characteristics of remote sensing satellite images; further, a jump-connected pixel reorganization and amplification module is introduced, which speeds up the calculation speed without a series of additional convolutions, and is well applied to power grid remote sensing image processing. It can be used for subsequent transmission line inspection, transmission line tower identification, insulator detection and other purposes.

The invention discloses an aerial photography small target detection method and system based on a deep neural network. The method comprises the following steps: acquiring aerial photography training set data; inputting the aerial photography training set data into a pre-constructed deep neural network for training, and generating an aerial photography small target detection model; obtaining a to-be-detected aerial photo; and inputting a to-be-detected aerial photo into the aerial photo small target detection model, and outputting a target classification result. According to the invention, through the aerial photography small target detection model, the target features in the aerial photography picture of the unmanned aerial vehicle are better extracted, and the detection precision of the small target is improved.

The invention discloses a distorted image correction method based on a fast multi-scale residual network. Relates to the field of image geometric distortion correction, in particular to the technical field of distortion image correction based on a fast multi-scale residual network. Comprising the following steps: S1, constructing a data set, and generating a paired data set of a distorted image and a standard image and a deformation displacement field file of a corresponding number of the distorted image by using a geometric distortion model; s2, a distorted image correction model based on a fast multi-scale residual network is constructed, the model is divided into a backbone network and a tributary network, the backbone network is used for recovering image features, and the tributary network is used for outputting image classification features; s3, training the distorted image correction model based on the fast multi-scale residual network by using the pairwise data set of the distorted image and the standard image and the deformation displacement field file; and S4, processing the to-be-corrected image by using the trained distortion image correction model based on the fast multi-scale residual network to obtain the corrected image and the corresponding classification thereof.

The invention discloses an artificial intelligence-based traditional Chinese medicinal material efficacy evaluation method and device, which can fully mine internal associations and modes in data based on artificial intelligent processing of large-scale multi-dimensional data, and finally improve the accuracy and stability of traditional Chinese medicinal material efficacy evaluation. The traditional Chinese medicinal material efficacy evaluation method comprises the following steps: performing data acquisition and data annotation in a traditional Chinese medicinal material production base, and storing the acquired data attributes in a structured JSON format; expanding the data by adopting a multiple linear interpolation algorithm based on a feature coupling degree, and generating a new data point by evaluating the coupling degree between input data features; performing feature extraction on the expanded data through a six-layer full-connection neural network; carrying out dimension reduction processing on the extracted features by adopting an auto-encoder based on reverse reconstruction; classifying the data after dimension reduction by using a support vector machine algorithm based on quantum state mapping; and processing a new traditional Chinese medicinal material sample by using the trained model so as to evaluate the efficacy of the traditional Chinese medicinal material.

A blind source separation method for a multi-source fault map of an oil-immersed power transformer comprises the following steps: inputting a multi-source partial discharge fault map into a CNN encoder, adding a ResNet network in a generative network, and introducing a gradient penalty term on the basis of a traditional CGAN network antagonism loss function; a deep learning dedicated server is used for multiple times of training and verification, model training parameters are continuously optimized and adjusted, and a network weight data file meeting oil-immersed power transformer partial discharge fault diagnosis task requirements is obtained; a trained YOLOv8 network is adopted as a diagnosis network, the fault type of each generated single-source atlas is identified, the fault type of a multi-source atlas is determined, and the fault state of the oil-immersed power transformer is deduced and diagnosed; and deploying the obtained Encoder-CGAN network and YOLOv8 network weight data files to power inspection equipment and a power inspection platform to complete fault monitoring and deduction diagnosis tasks of the oil-immersed power transformer. And the problem of deducing and diagnosing the fault type of the partial discharge fault map of the oil-immersed power transformer is solved.

The invention provides a vehicle driving obstacle distance detection method and system based on a vehicle-mounted depth camera, and relates to the technical field of vehicle obstacle detection. Comprising the following steps: acquiring RGB image information, depth point cloud information and 3D coordinate information of a to-be-identified obstacle during vehicle driving; geometric features of key points in the RGB image information are extracted, and obstacles in the key points are recognized; aligning and integrating the 3D coordinate information and the depth point cloud information according to a timestamp, and combining the identified obstacle to obtain a 3D coordinate with the timestamp of the obstacle; and learning spatial features of different scales of the 3D coordinates with the timestamps of the obstacle, and performing feature extraction to obtain the distance between the obstacle and the vehicle body. According to the method, the RGB image information, the depth point cloud information and the 3D coordinate information of the to-be-recognized obstacle in the running process of the vehicle are acquired by using the depth camera installed on the vehicle body, and classification and regression are performed after different feature information is extracted by using the hybrid network model, so that the effects of efficiently and accurately recognizing the obstacle and predicting the distance between the obstacle and the vehicle body can be considered.

The invention discloses a nuclear power material purchase demand prediction method and system based on a time convolutional network, and relates to the technical field of time series data analysis, and the method comprises the steps: collecting a data set, carrying out the preprocessing of the data set, classifying the data set, setting a representation mode of the input data set, and constructing a time convolutional neural network model. A data set is divided, model training is performed by using the divided data set, time sequence data is input into a trained model, and a material purchasing demand prediction value is output through multi-layer convolution, residual connection and full connection layer processing. According to the method, the time sequence data features are deeply learned through expansion convolution and residual connection, the ability of capturing the complex time dependency relationship by the model is improved, gradient flow in a deep network is promoted through introduction of the residual connection, the gradient disappearance problem is avoided, and the training efficiency and the learning ability of the complex features of the model are improved.

The invention discloses a digital display meter identification method and system based on an E-Net model and a LeNet-5 network. The method comprises the following steps: applying a bottleneck convolution structure to an ENet network model, and carrying out digital region segmentation on a to-be-identified image based on the E-Net model; carrying out numeric character extraction on the segmented numeric region by utilizing a binarization method of an adaptive threshold; and carrying out denoising and tilt correction on the extracted characters, and carrying out character recognition by using an attention mechanism-based LeNet-5 network model to obtain a digital display meter recognition result. According to the scheme of the invention, high-precision identification of the reading of the digital electric meter in various complex environments is realized.

The invention discloses a neural network training method and device based on depth time alignment gradient enhancement. The training method comprises the following steps: acquiring input sample data; inputting the input sample data into a neural network model based on depth time alignment gradient enhancement to obtain prediction category probability data corresponding to the category of the input sample data; according to the prediction category probability data and the sample label data, adjusting each parameter of the neural network model based on depth time alignment gradient enhancement to obtain a trained neural network model, the neural network model based on depth time alignment gradient enhancement comprising an input layer, K stages and an output layer, the first stage to the (K-1) th stage each comprise a convolutional network based on depth time alignment gradient enhancement and an auxiliary classifier network, the Kth stage comprises a convolutional network based on depth time alignment gradient enhancement, and K is a positive integer greater than 1.

The invention provides an LSTM network model for enhancing short-term memory, and belongs to the technical field of network models.The LSTM network model comprises a data input module, a data calculation unit, a feature extraction module and an output module, the data input module is connected with the data input end of the data calculation unit, and the data input module is used for inputting external processing data; the output end of the data calculation unit is connected with the feature extraction module and the output module, the output end of the feature extraction module is connected with the input end of the data calculation unit, the data calculation unit is used for memorizing and forgetting information, and the feature extraction module extracts feature values from vectors output by the data calculation unit and outputs the feature values to the output module. And then the data is circulated back to the data calculation unit, so that the processed data can be memorized again, and the data processing efficiency is improved.

The invention relates to the technical field of power grids, in particular to an electric power system stability evaluation method and system based on spatial-temporal characteristics. Acquiring first evaluation data of the power system, wherein the first evaluation data comprises bus voltage amplitude and phase angle data; inputting the first evaluation data into a preset RNN-AGCN evaluation model; and evaluating the stability of the power system based on an output result of the RNN-AGCN evaluation model. The first evaluation data comprises steady-state data before the fault, data in the fault and data after the fault is cleared, and the first evaluation data are collected through the PMU. Before data are input, min-max normalization processing is carried out on the data, and 1 * 1 convolution processing is carried out on the data to a specified dimension. The RNN-AGCN evaluation model comprises a recurrent neural network RNN and an adaptive graph convolutional network AGCN which are respectively used for extracting time sequence characteristics of transient data and spatial structure characteristics of power grid topology. Through double extraction of fusion time sequence and spatial features, the accuracy and anti-noise capability of evaluation are improved, and meanwhile, the method has relatively high topological adaptability and provides powerful guarantee for safe and stable operation of a power grid.

The present invention relates to the technical field of packaging product detection, and specifically to a quality detection method for packaging products based on the fusion of X-rays and visible light. In the present invention, X-ray images and visible light images are collected; feature engineering processing is performed on the X-ray images and visible light images to generate X-ray feature data and visible light feature data, wherein the data processed by feature engineering includes spatial and structural information; the X-ray feature data is used to detect the internal product quality using an internal quality detection model, and different output results are determined according to different internal product quality detection results; and the visible light feature data is enhanced according to the abnormal position to generate a visible light abnormal image, and the visible light abnormal image is used to detect the external product quality using an external quality detection model.

A K-CNN-BiGRU-based ship outfield coating atmospheric pollution concentration prediction method comprises the following steps: firstly, deploying an environment monitoring terminal to collect data such as weather and pollutant concentration, performing grid division on a research area, performing spatial interpolation by adopting IDW, performing calculation and arrangement to obtain a grid data set of atmospheric pollutant concentration, and calculating the atmospheric pollutant concentration according to the grid data set of the atmospheric pollutant concentration; the method comprises the following steps: firstly, calculating spatial correlation by using a KNN tree algorithm, screening to obtain a spatial correlation matrix, arranging and splicing the spatial correlation matrix into a space-time class diagram according to a time sequence, secondly, inputting the space-time class diagram into a CNN model to extract local space-time characteristics, and finally, sending the extracted space-time characteristics into BiGRU to carry out global prediction. According to the invention, the atmospheric pollution concentration of the ship outfield coating plant in the future period can be predicted, and employees are guided to scientifically carry out coating construction and safe and environment-friendly management.